On Friday, during Day 12 of its “12 Days of OpenAI,” CEO Sam Altman announced the latest AI “reasoning” models, o3 and o3-mini.
These models, which build on the o1 models released earlier this year, are not yet available for public use but are open for public safety testing and research access.
The new models employ what OpenAI calls a “private chain of thought,” allowing the AI to pause, reflect on its internal dialogue, and plan before responding. This method, known as simulated reasoning (SR), goes beyond traditional large language models (LLMs).
The model family is named “o3” instead of “o2” to avoid trademark issues with British telecom provider O2. During the announcement, Altman humorously acknowledged OpenAI’s naming challenges, stating, “In the grand tradition of OpenAI being, nasty at names, it’ll be called o3.”
The o3 model has achieved a record-breaking score on the ARC-AGI benchmark, a visual reasoning benchmark that has remained unbeaten since its creation in 2019. It scored 75.7 percent in low-compute scenarios and 87.5 percent in high-compute testing, comparable to human performance at an 85 percent threshold.
Additionally, o3 scored 96.7 percent on the 2024 American Invitational Mathematics Exam, missing only one question. It also achieved 87.7 percent on the GPQA Diamond benchmark, which includes graduate-level questions in biology, physics, and chemistry. On the Frontier Math benchmark by EpochAI, o3 solved 25.2 percent of problems, a significant improvement over previous models.
The o3-mini variant, also announced, includes an adaptive thinking time feature with low, medium, and high processing speeds, producing better results at higher compute settings. OpenAI reports that o3-mini outperforms its predecessor, o1, on the Codeforces benchmark.
The announcement comes as other companies develop their SR models, including Google’s Gemini 2.0 Flash Thinking Experimental, DeepSeek’s DeepSeek-R1, and Alibaba’s Qwen team’s QwQ, which they claim is the first “open” alternative to o1.
These new AI models are based on traditional LLMs but are fine-tuned for an iterative chain of thought processes that can scale at inference time, rather than focusing solely on model training improvements.
OpenAI plans to make the new SR models available to safety researchers for testing first. o3-mini is expected to launch in late January, and o3 will follow soon after.